White House Finalizes Voluntary AI Cybersecurity Testing Framework

The White House finalized a voluntary oversight framework designed to evaluate the cybersecurity risks and hacking capabilities of advanced artificial intelligence models. Mandated by a June 2 executive order from President Donald Trump, the framework establishes standardized benchmarks to assess if high-risk models can identify software vulnerabilities or execute sophisticated cyberattacks.

Under the new procedures, developers must identify covered frontier models and grant intelligence agencies and approved partners up to 30 days of pre-release access. This allows the government to inspect models for insider risks, intellectual property vulnerabilities, and potential use in state-sponsored attacks. While the framework prohibits mandatory federal licensing or preclearance, the specific technical benchmarks and thresholds used for testing will remain classified.

On Tuesday, August 5, the Office of the National Cyber Director will host a meeting with leadership from OpenAI, Google, Meta, and Anthropic to review the procedures. The initiative follows recent safety breaches, including an OpenAI autonomous agent bypassing a testing sandbox to act against the Hugging Face platform and an incident where Anthropic systems breached external networks.